基于SKNet注意力机制的飞机类型识别算法  

Aircraft type recognition algorithm based on SKNet attention mechanism

作　　者：舒振宇[1] 秦昊 SHU Zhenyu;QIN Hao(College of Electronics and Information Engineering,South-Central Minzu University,Wuhan 430074,China)

机构地区：[1]中南民族大学电子信息工程学院,武汉430074

出　　处：《中南民族大学学报（自然科学版）》2024年第1期69-77,共9页Journal of South-Central University for Nationalities：Natural Science Edition

基　　金：中南民族大学横向科研基金资助项目(HZY18017)。

摘　　要：飞机类型识别是细粒度图像分类的一种,重点在于设计神经网络模型使其能够分辨各飞机种类中细微而具有区分性的特征.针对当前飞机识别任务中飞机种类众多、类间差异细微、类内差异显著等问题,提出了一种基于改进SKNet注意力与数据增广的飞机类型识别算法.以ResNeXt101网络作为基础网络,改进CBAM注意力提出并行的通道-空间注意力PCSA并嵌入可选择卷积模块的不同分支,得到PCSA-SK注意力,将其嵌入基础网络以进一步融合基础网络提取的深层特征并为其分配权重.根据目标激活图中具有判别性信息的区域,在原图像上对判别性区域裁剪并加入训练集,实现数据增广.实验结果表明:该算法在FGVC-Aircraft数据集上取得了93.57%的识别准确率,优于其他飞机识别算法.Aircraft type recognition is a type of fine-grained image classification that focuses on designing neural network models capable of discerning subtle and distinctive features among various aircraft types.In response to challenges such as a large number of aircraft categories,subtle inter-class differences,and significant intra-class variations in aircraft recognition tasks,an aircraft type recognition algorithm is proposed based on improved SKNet(Selective Kernel Network)attention and data augmentation.The ResNeXt101 network utilized as the base network,improves the CBAM(Convolutional Block Attention Module)attention,and introduces a parallel channel-spatial attention called PCSA(Parallel Channel-Spatial Attention)embedded with different branches of selectable convolution modules,resulting in a new convolution module called PCSA-SK,which is integrated into the base network to further fuse and allocate weights to the deep-level features extracted by the base network.According to the region with discriminative information in the target activation map,the discriminative region is cropped on the original image and added to the training set to achieve data augmentation.Experimental results demonstrate that the proposed algorithm achieves a recognition accuracy of 93.57%on the FGVC-Aircraft dataset,outperforming other aircraft recognition algorithms.

关 键 词：飞机类型识别 SKNet注意力 数据增广 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]